mazurka project

Mazurka Project

29 June 2006

CHARM Symposium

Craig Stuart Sapp

Sections

1. Overview2. Power Measurements3. Manual Correction4. Automatic Alignment 5. Experiments6. Performance Simulations7. Initial Analysis

Top-Down Overview

Mazurka Project

Data Entry Data Analysis

Year 1 focusYear 2 focus

How to measure performance features fromaudio?

What to do with measuredfeatures?

Top-Down Overview

Mazurka Project

Data Entry Data Analysis

Score Preparation

Automatic Alignment Manual CorrectionAudio Preparation

Beat Tapping

Data Entry Process

Automatic Alignment Manual CorrectionBeat Tapping

• Tapping to beats in audio• Alignment of taps to audio times• Previously done with non-graphical program• Evaluating process in Sonic Visualiser

• tapping alignment to audio done automatically• time resolution not as accurate? (maybe 10 ms instead of 5 ms)

• Tapping accuracy is about 50/60 ms away from beat on average for mauzurkas, or 25 ms away from beat for a steady tempo.

Data Entry Process


• Refines taps by searching for score notes in neighborhood• Estimates and locates non-beat notes• Measures event amplitudes (loudness)

• Improves on tapping positions by 4-5x for modern recordings • 2-3x improvement for historic (noisy) recordings

Data Entry Process


• Fix obvious errors in automatic alignment

• Verify automatic alignment by listening to clicks and music simultaneously

Reverse Conducting• Orange = individual taps (multiple sessions) which create bands of time about 100 ms wide.

• Red = average time of individual taps for a particular beat

• Blue = actual beat onset times

Good localization, but not accurate average

Poor localization, but accurate average

Typically one beat delay in response

Power Measurements(for manual corrections)

MzPowerCurve• Sonic Visualiser plugin to do various power measurements

• #1 raw power measurements – average and weighted

• #2 smoothed power – useful for basic dynamics measurements

• #3,4 smoothed power slope – useful for manual corrections of note attacks (for percussive instruments such as piano).

http://sv.mazurka.org.uk/MzPowerCurve

Window and Hop

Overlapping In this example

Raw and Smoothed power

Note attack Note attack

(Measuring power every 10 ms, with a 10 ms window.)

• Smoothed power useful for getting basic dynamics levels.

Need to smooth to remove beatings and other quick artifacts.

Smoothing Filter• Using a filter called an exponential smoother:

Englishish: The current output equals the current input times the value k, plus the previous output times the value 1-k.

• All filters delay the input. Since this filter feeds back on itself, the filter will delay some frequencies more than others:

Delays for k = 0.3

Symmetric Filtering

Filtering Direction

Smoothing Direction• Avoid the funny delays by symmetric filtering

• Then slope of smoothed power aligns nicely with percussive note attacks

no

te a

ttack

• symmetric filtering is best for localizing attacks

• reverse filtering is best for dynamics estimation

Smoothed Power Slope

• smoothing factor of about 0.2 gives best results over a variety of conditions

Attack ½ between measurementsAttack on measurements

misses

Power Slope for Correcting

Not helpful for this beat (no peak)

Non-Synchronous Hands

70 ms separation

• Pianist is probably conveying a sense of relaxation at this point in music

• > 30 ms separation may be significant

Advantages/Disadvantages

• Time domain analysis, so localization can be better than for frequency analysis metrics (E.g. Earis & Bello methods)

• Ignores frequency content, so not always or accurate.

•Good for instruments with percussive attacks (i.e. piano, drums)

• Probably not good for non-percussive instruments: voice, violin, woodwinds, brass, etc.

Manual Correction

Beats + Offbeats

scaled power slope:

waveform:

beat beat beatoffbeat

Advantage of Having the Score

meaningless peaks(due to beating)

• If you don’t have a score, you are wasting your time.

Large-scale View of Beats/Offbeats

• Layer clicks can be played with different timbres/loudness.

Probable Entry Scenario0. Become familiar with the performance. (Score already entered)

• Tap to performance in Sonic Visualiser

• Cursory check of beat positions with onset annotations

• Interpolate off-beat positions based on score

• View/listen to audio with beats/off-beats and automatic annotations

• Automatic adjustments of the onset times of beats/off-beats

• Careful manual proof listening/reading of the automatically adjusted times

• Extract secondary performance features such as dynamics and non-simultaneous chord notes.

(10 min)

(5 min)

(10 min)

(30 min)

red: manual time estimates for a 5 minute piece

(15 min)

about 2 hours for 5 minutes of music

Automatic Alignment Evaluation

Summary

• Earis system parameter search optimization

1. wavenumber (k)2. low-pass filter order (LPF)3. tuning factor

• Other evaluation/exploration for the system:

1. search window method2. square/Gaussian window method3. recursive processing 4. wanderer identification 5. removing harmonics of previous event6. symmetric LPF filtering

• Automatic alignment improves accuracy about 4-5x for modern recordings and 2-3x for historic recordings when compared to reverse conducting accuracy.

k Parameter Sensitivity

k = wave number (how many cycles of a sinewave) to analyze with

LPF Parameter

Tuning parameter

Windowing Methods a = [ -0.33 ioi : 0.33 ioi ] b = [ -0.50 ioi : 0.50 ioi ] c = [ -0.67 ioi : 0.67 ioi ] d = [ -0.33 ioi : 0.67 ioi ] e = [ -1.00 ioi : 0.67 ioi ] f = [ -100 ms : 100 ms ] g = [ -200 ms : 200 ms ] h = [ -300 ms : 300 ms ] i = [ -400 ms : 400 ms ] j = [ -500 ms : 500 ms ] k = [ -200 ms : 500 ms ] l = [ -1 sd : 1 sd ] m = [ -2 sd : 2 sd ] n = [ -3 sd : 3 sd ] o = [ -4 sd : 4 sd ] p = [ -5 sd : 5 sd ] q = [ -2 sd : 4 sd ]

• IOI is best method

Hybrid Window Shape

all beats good beatsChiu 1999: 100% 60%Friedman 1930: 75% 50%

Recursion effect on localization

Recursion & Wanderingmodern recording

historic recording

Wandering

wandering

not wandering

Wanderers

Wanderers (2)

• events which the analysis method cannot “see”.

Wandering (3)

Selective Harmonics

• Removing shared harmonics with previous events didn’t help: more wanderers.

• Remove shared harmonics with previous event to improve attack identification and remove potentially beating harmonics.

Selective Harmonics (2)

Bello Onsets• Spectral measurements used to identify event onset locations• Can give false positivies and false negatives• Does not utilize a score

http://mazurka.org.uk/auto/onset

false negative false positive

LH/RH?

Onset Snapping

Chiu 1999, Mazurka 7/2: iteration 1 mean deviation: 10.5 ms bello snapping md: 13.4 ms

Friedman 1930, Mazurka 7/2: iteration 1 mean deviation: 22.0 ms bello snapping md: 22.1 ms

• Snap earis to bello if a bello onset is less than 50 ms away.

Stereo Differences

Stereo Comparison

Experiments

Note Onset Time Resolution

zoom in

rightchannel

leftchannel

• Play one of these clicks and ask listener: did it come before or after the start of the piano note?

piano note

Note Onset Time Resolution

• Very accurate to distinguish which came first when difference is > 60 ms.

• 75% accuracy or better outside -21.2 to +22.0 ms range around note attack.

• Symmetric about the note onset.

zone of confusion

Tapping Accuracy

• Tap for 30 minutes to a constant tempo

• 50% of taps occur within +/- 25 ms of actual event

• 95% of taps occur within +/- 50 ms of actual event

faster slower~25% ~25%

-50ms +50ms+25ms+25ms

• For Mazurkas (significant tempo changes), accuracy is about twice as much (50% occur within 50 ms of actual event).

Unpredictable Tempo Changes• Tapping to an unknown sudden change in tempo

stimulus:

• Suddenly faster:

responses:

average:

Same data as a tempo plot:

Unpredictable Tempo Changes (2)

• Tapping to an unknown sudden change in tempo

stimulus:

• Suddenly slower:

responses:

average:

Same data as a tempo plot:

Predictable Tempo Changes

• Tapping to an known sudden change in tempo:

Predictable Tempo Changes

Tempo JND

• How little can the tempo change before it is noticed?

60.0 63.33:

60.0 62.14:

60.0 62.05:

60.0 62.210:

61.05

60.67 61.33

60.28 60.55 60.83 61.10 61.38 61.65 61.93

Tempo JND (2)

Performance Simulations

Performance Feature Layers

constant

beat

event

notes

Tempo:

http://mazurka.org.uk/ana/midi

Performance Components

Tempo/Timing Dynamics

1. Average tempo (of entire piece)

2. Beat-to-beat tempo

3. Sub-beat timings (continuous tempo)

4. Non-simultaneous events (LH/RH, arpgggios)

1. Score dynamics

2. Composite loudness

3. LH/RH loudness

4. Individual note loudness

PerfViz

info(matchFileVersion,2.0).info(scoreFileName,'STDIN').info(midiFileName,'pid9048-06-corbsdpv.mid'). info(midiClockUnits,480). info(midiClockRate,500000). info(keySignature,[an,minor]). info(timeSignature,3/4). info(approximateTempo,102.4). snote(n1,[e,n],5,0:3,0,1,0,1,[])-note(1,[e,n],5,1656,2428,2428,43). snote(n2,[f,n],5,1:1,0,3/16,1,1.75,[])-note(2,[f,n],5,2428,2925,2925,49). snote(n3,[e,n],5,1:1,3/16,1/16,1.75,2,[])-note(3,[e,n],5,2925,3090,3090,41). snote(n4,[a,n],3,1:2,0,1,2,3,[])-note(4,[a,n],3,3090,3366,3366,40). snote(n5,[d,n],4,1:2,0,1,2,3,[])-note(5,[d,n],4,3090,3366,3366,40). snote(n6,[f,n],4,1:2,0,1,2,3,[])-note(6,[f,n],4,3090,3366,3366,40). snote(n7,[d,n],5,1:2,0,1,2,3,[])-note(7,[d,n],5,3090,3642,3642,40). snote(n8,[a,n],3,1:3,0,1,3,4,[])-note(8,[a,n],3,3642,3912,3912,43). snote(n9,[d,n],4,1:3,0,1,3,4,[])-note(9,[d,n],4,3642,3912,3912,43). snote(n10,[f,n],4,1:3,0,1,3,4,[])-note(10,[f,n],4,3642,3912,3912,43). snote(n11,[f,n],5,1:3,0,1,3,4,[])-note(11,[f,n],5,3642,4181,4181,39). snote(n12,[f,n],5,2:1,0,2,4,6,[])-note(12,[f,n],5,4181,5301,5301,62). snote(n13,[a,n],3,2:2,0,1,5,6,[])-note(13,[a,n],3,4649,4975,4975,39). snote(n14,[c,n],4,2:2,0,1,5,6,[])-note(14,[c,n],4,4649,4975,4975,39).

• 3D performance worm visualiser by Martin Gasser (Vienna)

MIDI file + Match file

PerfViz (2)

3 Axes:1. Time2. Tempo3. Loudness

Initial Analysis

Beat Durations

mazurka 2nd beat accentuation

long

er b

eat

shorter beat

Beat Durations (2)

Tempo and Style

Beat-duration plot

Rokudan no shirabe; Kin’ichi Nakanoshima (1904-1984)

JVC VICG-60397

Average tempo over time

Average tempo over time (2)

Mazurka in F Major, Op. 68, No. 3

Rubinstein 1938

Rubinstein 1961

Chiu 1999

MM134

MM129

MM115

Repeats

Dynamics

mazurka project

Documents